SenseNova-Loop-9B-Preview
modelYour notes
A SenseTime preview of a looped multimodal language model, released October 30, 2026 under Apache 2.0 for long-context and agentic work (text and images in, text out). Its 32 distinct decoder layers hold about 9B parameters, but the middle block of 24 layers (layers 4–27) runs twice with shared weights, so each forward pass executes 56 layers, about 14.6B parameters' worth of compute, while storing only 32. Inside the loop, two residual streams are joined by manifold-constrained hyper-connections (mHC): learned mappings, with the stream-mixing matrix Sinkhorn-normalized over 20 iterations, decide how each layer reads from and writes back to the two streams. Attention is hybrid, three linear-attention layers to every full-attention layer, with loop boundaries aligned to those four-layer groups. Training sequences reach 262,144 tokens and include text, images, code and tool trajectories. The card does not say whether the decoder was pretrained from scratch. Its 248,320-token vocabulary and three-linear-to-one-full layout match Qwen3.5's.
The card's benchmark chart covers agentic and office work (BIRD-Dev, WorkBuddy Bench, AIDABench, TeamBench, Terminal-Bench 2.1, T³-Banking, AutomationBench and SenseTime's own RaccoonPPTBench), and the showcase demos come from SenseTime's Raccoon Work agent. MTP speculative decoding ships now, while DSpark and DFlash2 draft heads are still training. The technical report and a Transformers implementation are listed as "TBD".